Research map / 3D vision / 3D generation
FlareX: A Physics-Informed Dataset for Lens Flare Removal via 2D Synthesis and 3D Rendering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds FlareX, a physics-informed lens flare removal dataset mixing 9,500 parameterized Blender flare templates synthesized into 2D images with illumination-law brightness adjustment, plus 3,000 3D-rendered pairs from 60 scenes, and a masking-based real-world test set.
Problem
Existing lens flare removal datasets are synthesized by overlaying limited 2D flare templates without physical principles, and real-world paired flare data with clean ground truth is hard to capture, so models generalize poorly to real flares.
Approach
Generate flare data in three stages: parameterized Blender flare templates, 2D synthesis with a brightness adjustment module based on the illumination law (using monocular depth and field-of-view-based incidence angle), and 3D Blender scene rendering with physically placed flares; test ground truth is obtained by occluding the light source rather than cleaning the lens.
Key results
- Uformer trained on FlareX reaches 25.459 PSNR, 0.692 SSIM, 0.133 LPIPS on the proposed test set, beating the second-best Uformer result (Flare7K++ at 24.079 PSNR, actually FlareReal600 at 24.079) by about 1.38 dB.
- On off-screen flare images, Uformer trained on FlareX achieves the lowest NIQE (3.967) and BRISQUE (28.354).
- Across HINet, MPRNet, Uformer, Restormer, and AST, models trained on FlareX outperform the same models trained on Wu et al., Flare7K, Flare7K++, and FlareReal600 datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train Uformer on FlareX instead of Flare7K++ | Uformer · FlareX test set | PSNR: nearly 1.38 dB | Helped |
| Train Uformer on FlareX instead of Flare7K++ | Uformer · FlareX test set | SSIM: 3.75% | Helped |
| Train Uformer on FlareX instead of Flare7K++ | Uformer · FlareX test set | LPIPS: 5% | Helped |
| Train the restoration models on FlareX instead of existing datasets | HINet, MPRNet, Uformer, and Restormer · FlareX test set | PSNR | Helped |
| Train Uformer on FlareX instead of other datasets | Uformer · Off-screen flare images | NIQE | Helped |
| Train Uformer on FlareX instead of other datasets | Uformer · Off-screen flare images | BRISQUE | Helped |
| Add illumination-law brightness adjustment to flare synthesis (ablation) | FlareX-trained Uformer · FlareX test set | model performance | Helped |
| Add illumination-law brightness adjustment to flare synthesis (ablation) | Flare7K-trained Uformer · Flare7K test set | model performance | Helped |
| Combine Flare-2D and Flare-3D instead of training on only part of FlareX (ablation) | Uformer and Restormer · FlareX test set | model performance | Helped |
Surprising findings
- Existing test-set ground truths retain strong and reflective flares, and nearly one-third of samples in existing test sets fail to reflect the Flare7K to Flare7K++ improvement in user study.
Limitations
- Extremely heavy flare remains challenging and leaves visible artifacts after restoration
- Estimated depth maps are not always accurate, motivating the 3D rendering stage
- Camera parameters are unknown, so incidence angle is estimated from an assumed horizontal field of view
Experimental setup
Datasets: Flare7K Flare7K++ FlareReal600 24K Flickr background images
Benchmarks: FlareX test set (100 real pairs) Flare7K test set FlareReal600 off-screen flare images (63 internet images)
Metrics: PSNR SSIM LPIPS NIQE BRISQUE user study
Baselines: Wu et al. dataset Flare7K Flare7K++ FlareReal600
Builds on: Flare7K Flare7K++ Blender Flared plugin Uformer
Closest papers by meaning
- Disentangle Nighttime Lens Flares: Self-supervised Generation-based Lens Flare Removal (AAAI 2025 · similarity 0.92)
- FLARE: Feed-forward Geometry, Appearance and Camera Estimation from Uncalibrated Sparse Views (CVPR 2025 · similarity 0.86)
- Language-Guided One-Step Diffusion Model for Nighttime Flare Removal (CVPR 2026 · similarity 0.86)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.84)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.83)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.